The Research Of NART2 Network Group And Its Application

ART2 network of the adaptive resonance theory system is an unsupervised competition neural network structure, that realizes classification through competition and self-stabilization mechanism, and can process the arbitrary sequence binary and analog input sample in real-time. ART2 network is not only able to identify learned-samples quickly, but also can quickly identify the new sample. Just owing to its fast learning rule, the neural network exists Stability-Plasticity trade-off trouble which is measured by the reaction to new input sample. When network stability is very strong, the network is less vulnerable to influence by new input sample and the LTM vector of long memory system is less vulnerable to change too, however the plasticity of network is opposite.Basing on the above problems, and combining with the neoteny characteristic of evolutionary biology, this paper proposes an improved algorithm-the NART2 network with the neoteny characteristic. The network can show strong plasticity in initial stage of learning sample, and show strong stability in the terminal stage. In addition, NART2 network can effectively restrain the influence caused by input sequence of the sample. In order to enhance the recognition ability and the flexibility of ART2, integrating schema theory of cognitive psychology, this paper designs vigilance value adjustment mechanism with teacher’s supervision. The vigilance value adjustment mechanism overcomes the defect which the vigilance value is fixed in the traditional ART2 network. NART2 with vigilance value adjustment mechanism is more consistent with human cognitive process. The experiment demonstrates the classification performance of NART2 network with the vigilance value adjustment mechanism is superior to conventional ART2 network.In order to improve NART2’s the capability of generalization and varied processing for data types, this paper presents a new Neural Network Group NART2 network group. The network group with accordant detection characteristic can filter characteristic data that could not stand for feature of samples, so as to increase accuracy of classification and simplify calculation. The vote mechanism is used into vigilance value adjustment mechanism of NART2 network group. After achieving vote, by adjusting vigilance value and exerting forced learning function NART2 network group can effectively reduce the network fault rate. At the same time, when sample only includes partly characteristic of input vector, NART2 network group introduces mechanism of association and forgetting, which can give the correct outcome of input samples and obtain the other characteristics of input sample. Finally, neural network group is applied successfully into the fruit’s image recognition, and identification result is quite wonderful.